Asri, M. (BMEB) – Genome Inference in the Pangenome Era

Human genome references have provided unified coordinate systems for mapping genetic variation, enabling numerous discoveries in genomics. However, mapping to a single linear reference can be misleading or even impossible when a sample’s genome is highly divergent from the reference, an issue known as reference bias. Recent advances in long-read sequencing technologies and computational techniques have made generating high-quality genome assemblies at scale more feasible than ever. This shift has paved the way for efforts such as the Human Pangenome Reference Consortium (HPRC), whose mission is to build a human pangenome reference that represents human genetic diversity and mitigates reference bias. In this dissertation, I present methods I developed that contribute to building more accurate pangenome references and advancing their applications in variant calling, ultimately enabling more accurate genome inference.

First, I introduce Secphase, a tool for correcting read mappings with direct applications in fixing assembly errors. Secphase corrects mismapped Pacific Biosciences (PacBio) HiFi reads by leveraging long-range phasing information from Oxford Nanopore Technologies (ONT) data. As a core component of the PHARAOH pipeline, Secphase was coupled with DeepPolisher, a transformer-based polisher, to polish 230 HPRC assemblies, reducing errors by half.

Second, I present HMM-Flagger, a reference-free tool for detecting structural misassemblies based on the coverage of mapped long reads. It models read coverage using a hidden Markov model (HMM), enabling it to classify coverage anomalies as erroneous blocks, false duplications, or collapsed blocks. HMM-Flagger has been used extensively in the evaluation and selection of genome assemblies for the HPRC releases.

Finally, as an application of pangenomes, I highlight work on extending DeepVariant to build a pangenome-aware variant caller by augmenting sample-specific input data with pangenome alignments. Using Element Biosciences reads, pangenome-aware DeepVariant generated call sets with up to 46.5% fewer errors than existing methods.

Event Host: Mobin Asri, PhD Candidate, Biomolecular Engineering & Bioinformatics 

Advisor: Benedict Paten

Zoom: https://ucsc.zoom.us/j/91210219698?pwd=HA0b0by9HcKlKiBua7anVBZ7ZfSfJu.1

Passcode: 594980